Abstract
Carbohydrate-protein supplementation often improves endurance performance. However, effectiveness varies significantly among individuals due to unique personal characteristics. This study aimed to develop a predictive machine learning framework for personalized supplementation, with a core methodological novelty in applying a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) to address the critical issue of data scarcity. Based on 231 rowing trials, the framework utilized 46 input features covering baseline characteristics and dietary intakes. Rowing distance was the performance outcome. The machine learning pipeline first utilized a hybrid feature selection method (correlation analysis, model-based importance, and domain knowledge). Following a comparative evaluation, WGAN-GP was utilized for data augmentation. Finally, several regression models (XGBoost, SVR, and MLP) were trained to predict rowing performance. The top-performing model was used to construct a personalized supplement recommendation framework. Feature selection identified 21 key indicators from 46 initial inputs. The XGBoost model, enhanced with WGAN-GP data augmentation, demonstrated the most robust performance, achieving a strong predictive accuracy (R² = 0.53) coupled with high stability. Body weight, explosive power, and nutritional inputs were key performance predictors. This study demonstrates that a data-augmented machine learning approach can effectively model individual responses to supplementation. The developed framework provides a data-driven pathway for creating personalized nutritional strategies to optimize athletic performance.